# Preprint reports tool-call repair gains without retraining AI model

_Published Monday, October 19, 2026 at 11:20 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers report that SchemaShield-Lite, a repair layer for AI agents calling software tools, improved accuracy when tool interfaces changed without changes to their underlying functions. In a preprint, tests on Qwen3.5-9B raised DICE-300 accuracy from 0.717 to 0.813 and BFCL-200 accuracy from 0.580 to 0.805.

The method validates argument schemas, retries when tool names are missing and makes one constrained repair attempt without retraining the model. It recovered 19/22 and 33/45 failures caused by interface changes, with no observed repair harms. Changes to the available tool list, especially distracting tools, caused most observed damage; the reported results cover one model and two benchmark evaluations.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2610.23342)

---
Canonical: https://techandbusiness.org/newswire/GGObDh4lkxS8hSR2YeMKz3
Published: 2026-10-19T15:20:34.580Z
Story chronology: 2026-10-19T04:00:00.000Z
Retrieved: 2026-10-19T18:28:57.234Z
Publisher: Tech & Business (techandbusiness.org)